主流大语言模型在输液港健康科普中的表现 评估：可读性与质量分析

Evaluation of Mainstream Large Language Models in Health Education on Infusion Port Care: Readability and Quality Analysis

摘要: 目的：系统评估主流大型语言模型(LLMs)生成植入式静脉输液港(IVAPs)科普文本的可读性与专业质量，为临床健康教育工具的选择提供循证依据。方法：选取5款主流通用大模型(GPT-5、豆包、深度求索、通义千问、文心一言)，针对输液港5大核心主题生成100篇科普文本。采用7项国际通用可读性指标，结合中文版患者教育材料评估工具(C-PEMAT-P)和全球质量量表(GQS)，对文本进行多维度量化分析。结果：不同模型生成文本的专业质量存在极显著差异(P < 0.001)，GPT-5综合表现最优，豆包和深度求索紧随其后；各模型文本的可读性同样存在显著差异。科普主题的复杂度会显著影响文本可读性，但不会改变高质量模型的内容产出水准，证实可读性与质量是两个相对独立的评价维度。结论：模型类型是决定输液港科普文本质量的核心因素，临床应用优先推荐GPT-5、豆包和深度求索。医护人员应优选高质量模型生成专业内容，再结合患者认知特点进行针对性的可读性优化，实现医学专业性与大众易懂性的平衡。

Abstract: Objective: To systematically evaluate the readability and professional quality of health education texts on implantable venous access ports (IVAPs) generated by mainstream large language models (LLMs), thereby providing an evidence-based basis for selecting clinical health education tools. Methods: A total of 100 texts were generated by 5 widely available LLMs (GPT-5, Doubao, DeepSeek, Tongyi Qianwen, Wenxin Yiyan), covering 5 core education themes related to IVAPs. Seven internationally validated readability indices, the Chinese version of the Patient Education Materials Assessment Tool (C-PEMAT-P), and the Global Quality Scale (GQS) were used for multidimensional quantitative analysis of the texts. Results: Significant differences in text quality were found among the models (P < 0.001), with GPT-5 achieving the highest overall performance, followed by Doubao and DeepSeek; significant differences in text readability were also observed across models. Topic complexity significantly affected readability but not the quality of content produced by high-performing models. Topic complexity significantly affected readability but not quality, indicating that readability and quality are relatively independent evaluation dimensions. Conclusion: Model type is the key determinant of text quality on IVAPs, with GPT-5, Doubao, and DeepSeek being the optimal choices. Healthcare providers should prioritize high-quality models for professional content generation, then optimize readability based on patients’ cognitive characteristics to balance medical accuracy and public comprehensibility.